SAP generativeAI — Connecting AI to SAP Processes
Generative AI in SAP landscapes rarely fails on the model. It fails on the connection: where the data comes from, who is allowed to see it, and what happens when the model is wrong. Those are interface questions, and interfaces are what we have done for over twenty years.
AI Core and model connectivity
SAP AI Core is where models are run and reached from SAP applications. The interesting part is not the call itself but everything around it: which system triggers it, what the fallback is when the model is unavailable, and how the response gets back into a business process in a usable shape.
RAG on enterprise data
Retrieval on your own documents and master data is what turns a general model into something useful in your organisation. It is also where the authorisation question becomes unavoidable: retrieval has to respect what the asking user is cleared to see, otherwise the assistant leaks across boundaries that the applications themselves enforce.
Governance that survives an audit
What was asked, what was answered, and on which sources — if that cannot be reconstructed, the use case will not clear an internal review. Logging and traceability are cheap to build in at the start and expensive to retrofit.
Frequently Asked Questions
Does our data go to a public model provider?
Only if you design it that way. SAP AI Core supports models under different hosting arrangements, and which one applies is a deliberate decision with contractual and data protection consequences. It should be settled before the first prototype, not after.
What is a realistic first use case?
Something narrow, with a clear owner and a measurable before-and-after. Document classification, drafting standard correspondence or answering questions on internal rulebooks all work. Anything that has to be right every single time is a poor first choice.
How does this relate to SAP Joule?
Joule is embedded in the SAP applications and knows their context, but is limited to what SAP provides there. A custom solution on AI Core is freer and has to establish that context itself. In practice the two complement each other rather than compete.
What does it cost to run?
Cost scales with usage, not with the licence. The drivers are the number of calls, the size of the context sent with each one, and how much retrieval happens per question. A design that sends everything to the model on every request gets expensive quickly and quietly.